# NYSED ↔ NYC DOE Chronic-Absenteeism Reconciliation

Date: 2026-06-09. First cross-publisher reconciliation of NYSED Report Card chronic
absenteeism (CA) against the NYC DOE end-of-year per-grade CA data in the local Postgres
replica (`edu_data_local.school_year_metrics`). School years compared: **2018-19, 2023-24,
2024-25**. Expanded 2026-06-10 to **2020-21, 2021-22, 2022-23** — see §9.

## 1. Method

### NYSED side
- Source zips (downloaded 2026-06-09): `SRC2019`, `SRC2022`, `SRC2024`, `SRC2025` from
  data.nysed.gov. Extracted the `.mdb` from each, exported tables
  `ACC EM Chronic Absenteeism` (grades 1–8) and `ACC HS Chronic Absenteeism` (grades 9–12)
  with `mdb-export`, then deleted the mdbs. CSVs in
  `nyc-edu-data/data/reference/sources/nysed_src/2026-06-09/extracted/` (gitignored).
- YEAR encoding confirmed from SRC2025 ReadMe: "YEAR … 2024 for 2023-24, 2025 for 2024-25".
  So SRC2019 YEAR=2019 → 2018-19 SY; SRC2024 YEAR=2024 → 2023-24; SRC2025 YEAR=2025 → 2024-25.
- Rows used: `SUBGROUP_NAME='All Students'`, NYC (ENTITY_CD county prefix 31–35),
  school-level (digits 9–12 ≠ '0000'), **district schools only** (digits 7–8 ≠ '86').
  ABSENT_RATE parsed as a number; `s` (suppressed) rows skipped.
- Revision check: SRC2024 and SRC2025 YEAR=2024 NYC district-school All-Students rows are
  **byte-identical** (EM 1166/1166, HS 505/505) — no inter-release revisions; own-year file
  precedence is moot for these rows.

### Our side (DOE)
- `school_year_metrics`, `metric_key = chronic_absenteeism_rate_grade_1 … _grade_12`,
  `subgroup='ALL'`, `not suppressed and value is not null and denominator > 0`. Single
  provenance: data_load `644077af` "NYCDOE — End-of-Year Attendance + Chronic Absenteeism
  (2018-19 to 2024-25)" (one xlsx covering all years uniformly).
- Band aggregate per school-year: `N = Σ denominator`, `absent_est = Σ(value/100 ×
  denominator)`, `rate = 100 × absent_est / N`, grades 1–8 (EM) and 9–12 (HS).
- Joined on (derived DBN, school year). District DBNs only (`^(0[1-9]|[12]\d|3[0-2])[MXKQR]\d{3}$`).

### File-format findings (step 2)
- All columns export as quoted text. ABSENT_RATE is one decimal place (`48.6`) or a bare
  integer (`39`); no `%` signs, no padding. ENROLLMENT/ABSENT_COUNT are bare integers.
  Suppression marker is the literal `s` in the value columns. The only whitespace issue is
  trailing spaces in some `ENTITY_NAME`s (84–1,910 rows/file); no whitespace in numeric fields.
- Layout variants as documented in `nysed-scout/FORMAT_NOTES.md`: SRC2019 has the MIP-era goal
  columns and no INSTITUTION_ID; SRC2025 drops LEVEL/OVERRIDE.
- Row counts: SRC2019 EM 81,731 / HS 37,848; SRC2022 EM 87,057 / HS 40,900; SRC2024 EM
  92,270 / HS 43,440; SRC2025 EM 92,050 / HS 43,410. Each file carries exactly two YEAR
  values (current + prior).

## 2. BEDS → DBN derivation

Rule applied: borough letter from county code (31→M, 32→X, 33→K, 34→Q, 35→R); 2-digit
district = digits 3–4; DBN number = **last 3 digits** of ENTITY_CD. Verified structure:
digits 5–8 are `0001` for every NYC district-school row in all six panels.

- **Digit 9 is not always '0'.** 1,658 rows (465 distinct schools across years/tables) have
  digit 9 = '1' — these are buildings numbered 1xxx (e.g. `331300011430` → 13K430 Brooklyn
  Tech, `310100011450` → 01M450 East Side Community). No other digit-9 values occur. The
  last-3-digits rule still produces the correct DBN.
- **Zero DBN collisions** in any (year, table) panel — no 0xxx/1xxx building-number ambiguity
  materialized.
- **D75: not derivable and moot — NYSED CA tables contain no D75 schools at all** (0 rows
  with ENTITY_CD prefix `3075` in all 8 exports; the FORMAT_NOTES D75 examples exist as
  institutions but have no CA rows). D79 also absent (0 rows prefix `3079`), confirming the
  scout finding. Headline stats are district schools only, as planned.
- Charters (digits 7–8 = '86') excluded from the join (no deterministic DBN; crosswalk needed).

## 3. NYC coverage per panel

| Panel | NYC schools | charters | district | district w/ rate | suppressed `s` | DATA_REP_FLAG=N |
|---|---|---|---|---|---|---|
| 2018-19 EM (SRC2019 Y=2019) | 1,354 | 194 | 1,160 | 1,120 | 40 | 0 |
| 2018-19 HS | 621 | 72 | 549 | 466 | 83 | 0 |
| 2023-24 EM (SRC2024 Y=2024) | 1,409 | 243 | 1,166 | 1,132 | 34 | 49 |
| 2023-24 HS | 598 | 93 | 505 | 469 | 36 | 5 |
| 2024-25 EM (SRC2025 Y=2025) | 1,411 | 250 | 1,161 | 1,132 | 29 | 45 |
| 2024-25 HS | 603 | 96 | 507 | 471 | 36 | 6 |

### Success Academy non-reporting (DATA_REP_FLAG = 'N')
SA stopped reporting attendance to NYSED network-wide starting 2023-24:

| Year | EM: SA flagged N / SA present (of all NYC N) | HS |
|---|---|---|
| 2018-19 | 0 / 28 (0 total N) | 0 / 3 |
| 2021-22 (SRC2022) | 0 / 29 (4 total N) | 0 / 3 |
| 2023-24 | **31 / 31** (of 49 total N) | **3 / 3** (of 5) |
| 2024-25 | **35 / 35** (of 45 total N) | **3 / 3** (of 6) |

SA is the bulk of the non-reporting universe (charters only; doesn't affect the
district-school reconciliation below).

## 4. Reconciliation results (district schools, All Students)

diff = our rate − NYSED rate, percentage points. Enrollment ratio = our N / NYSED ENROLLMENT.

| Panel | joined | NYSED-only | ours-only | mean | median | SD | p10 | p90 | ≤0.5pp | ≤1.0pp | ≤1.5pp | corr | enr ratio med [p10,p90] |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 2018-19 EM | 1,118 | 2 | 13 | +0.58 | +0.39 | 1.32 | −0.33 | +1.75 | 50% | 74% | 86% | 0.995 | 1.017 [1.000, 1.094] |
| 2018-19 HS | 414 | 52 | 3 | +0.32 | +0.09 | 1.16 | −0.74 | +1.75 | 52% | 75% | 86% | 0.997 | 0.993 [0.974, 1.000] |
| 2023-24 EM | 1,119 | 13 | 5 | +0.48 | +0.38 | 1.08 | −0.45 | +1.62 | 48% | 74% | 85% | 0.997 | 1.033 [1.003, 1.112] |
| 2023-24 HS | 410 | 59 | 6 | +0.34 | +0.30 | 0.88 | −0.39 | +1.16 | 53% | 80% | 90% | 0.998 | 1.000 [1.000, 1.003] |
| **2024-25 EM** | 1,115 | 17 | 5 | **−1.74** | **−1.84** | 1.57 | −3.37 | +0.19 | **11%** | **22%** | **35%** | 0.993 | 1.027 [1.000, 1.111] |
| 2024-25 HS | 410 | 61 | 7 | −0.20 | −0.02 | 1.11 | −0.70 | +0.12 | 84% | 89% | 91% | 0.997 | 1.000 [0.999, 1.002] |

Enrollment-weighted aggregates over the joined set:

| Panel | NYSED (count/enr = rate) | Ours (est/N = rate) | diff |
|---|---|---|---|
| 2018-19 EM | 115,749/559,530 = 20.69% | 119,118/562,288 = 21.18% | +0.50pp |
| 2018-19 HS | 89,790/284,064 = 31.61% | 88,027/277,796 = 31.69% | +0.08pp |
| 2023-24 EM | 140,903/469,427 = 30.02% | 145,635/477,711 = 30.49% | +0.47pp |
| 2023-24 HS | 95,513/270,198 = 35.35% | 94,644/265,579 = 35.64% | +0.29pp |
| **2024-25 EM** | 140,934/458,347 = 30.75% | 135,244/465,794 = 29.04% | **−1.71pp** |
| 2024-25 HS | 94,188/269,358 = 34.97% | 92,189/266,634 = 34.58% | −0.39pp |

Exact agreement: at HS, 71% of joined schools have **byte-identical denominators** in
2023-24/2024-25 (most HS serve only 9–12, so the universes coincide); in 2024-25 HS, 52% of
schools match the rate within 0.05pp. At EM, denominators rarely match exactly (ours ~1.7–3.3%
larger at the median), consistent with universe-rule differences (NYSED EM counts ungraded
age-equivalent students; window/inclusion rules differ).

### Join coverage notes
- **NYSED-only HS (~52–61/yr)** are overwhelmingly transfer/alternative high schools
  (e.g. 01M458 Forsyth Satellite, 01M515 Lower East Side Prep, 01M650 Cascades) plus a few
  6-12 schools (03M610/04M610 Young Women's Leadership). Checked in DB: these DBNs exist in
  `schools` but have **zero** metrics rows — the DOE EOY attendance file simply omits them.
  This is a coverage gap on our side, not a join failure.
- **Ours-only (3–13/panel)** are schools whose NYSED row is suppressed (`s`) or absent that
  year (e.g. 08X583, 03M291 West End Secondary 2018-19 HS — school too new/small for NYSED
  accountability reporting).
- NYSED-only EM examples 2023-24/2024-25: 03M610, 05M371 SEED Harlem (state-authorized
  boarding school, not in DOE file), 07X642 South Bronx Literacy Academy.

### Worst |diff| cases (top of each panel; full top-10 in `recon_results.json`)
- 2018-19 EM: 23K697 Teachers Preparatory HS, NYSED 51.6 vs ours 82.35 (N 62 vs 17) — see
  hand case 3: partial-grade artifact, not a data conflict.
- 2023-24 EM: 13K266 (38.6 vs 44.8, N 70 vs 87); 05M194 (48.7 vs 54.7, N 117 vs 137) —
  small schools where universe differences swing the rate.
- 2024-25 EM: 02M347 ASL & English Lower (49.1 vs 38.4, N 55 vs 125); 32K116 (36.2 vs 28.2);
  nearly all top-10 share the same sign (ours lower).
- 2024-25 HS: 25Q425 John Bowne (42.8 vs 34.1, N 2894 vs 2927, −8.7pp) — largest single-school
  conflict found; denominators nearly identical, numerator diverges.

## 5. Hand-verified cases (end-to-end arithmetic)

**Case 1 — 01M015 PS 15 Roberto Clemente, 2023-24 EM.**
NYSED (SRC2024, YEAR=2024): ENROLLMENT 142, ABSENT_COUNT 69, ABSENT_RATE 48.6.
Ours (grades 1–5): 35×45.714287% = 16; 29×51.724136% = 15; 30×53.333332% = 16;
33×42.424244% = 14; 23×47.826088% = 11 → 72 absent / 150 = **48.00%** vs NYSED **48.6%**
(diff −0.60pp; N 150 vs 142, ratio 1.056). Reconstructed counts are integers — the DOE rates
are exact count ratios.

**Case 2 — 13K430 Brooklyn Technical HS, 2024-25 HS.**
NYSED (SRC2025, YEAR=2025): 5,840 / 979 / 16.8.
Ours: g9 1,399×5.861330% = 82; g10 1,410×8.368794% = 118; g11 1,468×14.032698% = 206;
g12 1,563×27.063339% = 423 → 829 / 5,840 = **14.20%** vs NYSED **16.8%**.
**Denominator identical (5,840 = 5,840); numerator differs by 150 students.** DOE's own
published school-total CA for 13K430 2024-25 is 14.195% — matches our reconstruction exactly.
Same school in 2018-19: 10.2 vs 10.31 (+0.11pp); 2023-24: 14.0 vs 14.21 (+0.21pp). The
divergence is new in 2024-25 and numerator-only.

**Case 3 — 23K697 Teachers Preparatory HS (6-12 school), 2018-19.**
NYSED EM: 62 / 32 / 51.6. Ours: only grade 8 is non-suppressed (17×82.352943% = 14) →
82.35% on N=17. Grades 6–7 are missing from the DOE per-grade file, so our EM aggregate
covers 17 of ~62 students — the worst-case |diff| of +30.75pp is a **partial-grade-coverage
artifact**, not a publisher conflict. The HS side of the same school reconciles fine:
NYSED 217/122/56.2 vs ours 121/212 = 57.08% (+0.88pp).

## 6. The 2024-25 EM divergence (verified facts vs open cause)

Verified:
1. Five of six panels agree tightly (median |diff| ≤ 0.4pp, weighted aggregate within
   ±0.5pp, correlation ≥ 0.993). 2024-25 EM does not: median −1.84pp, weighted −1.71pp,
   only 22% of schools within ±1pp.
2. The divergence is **broad-based**, not compositional: median diff −1.92pp for
   elementary-only (gr 1–5) schools, −1.82 for K-8, −1.49 for middle-only (6–8).
3. It is **numerator-side, not denominator-side**: enrollment ratios are unchanged from
   2023-24 (median 1.027 vs 1.033); Brooklyn Tech shows identical N with a 150-student
   numerator gap; 59 of the 62 EM schools with byte-identical N still differ by >0.5pp
   (vs 2 of 22 in 2023-24).
4. The two publishers **disagree on the direction of the year-over-year change** for the
   same joined schools: NYSED EM CA 30.02% → 30.75% (up +0.7pp); DOE 30.49% → 29.04%
   (down −1.4pp). DOE's citywide per-grade rates fall in *every* grade PK–12 in 2024-25
   (−0.5 to −4.1pp); a milder echo exists at HS (ours −1.06pp YoY vs NYSED −0.38pp).
5. Not a NYSED revision or vintage issue: SRC2024 vs SRC2025 overlap rows are identical, and
   our DOE source is a single EOY file covering all years with the same layout.
6. Not a reconstruction artifact: for pure 6–8 schools, our grade-sum reproduces DOE's own
   published school-total rate with **median diff 0.000pp and identical N** in both years
   (199/335 and 197/335 exact; remainder explained by per-grade suppression/ungraded
   students).

**Cause not established from data on hand.** Candidate explanations to verify externally
(do not assert without sourcing): a DOE 2024-25 definitional change in the EOY file (the OSC
audit reportedly claims DOE excludes out-of-school-suspension and medical absences); a
NYSED-side change coinciding with CA's removal as an accountability indicator in 2025
(replaced by the Attendance Index; minimum-N dropped 30→20). The data show whatever changed
in 2024-25 removed ~5,700 students from the EM numerator (and ~150 at Brooklyn Tech alone)
relative to NYSED's count, while denominators track as before.

## 7. Caveats

- Our side reconstructs counts from published per-grade *rates* × denominators; this is exact
  where rates are exact count ratios (hand cases reproduce integers) but limited by per-grade
  suppression: schools with suppressed grades contribute partial aggregates (see Case 3).
  No minimum-coverage filter was applied (spec: any non-suppressed grades, N>0).
- Universe definitions differ by construction: NYSED ENROLLMENT = enrolled ≥10 instructional
  days and present ≥1 day, grades 1–8/9–12 **including ungraded age-equivalent students**;
  DOE per-grade denominators = students contributing 10+ days & 1+ present day, graded only.
  Median our/NYSED N ratio: ~1.02–1.03 EM, ~1.00 HS.
- NYSED suppression (`s`) removes 29–83 district schools per panel from the join; suppressed
  rows are excluded, biasing the joined set slightly toward larger schools.
- Transfer/alternative HS are absent from the DOE EOY file (not from NYSED) — HS join covers
  ~410 of ~470 rate-bearing NYSED district HS rows.
- Charters reconciled not at all here (no deterministic DBN; SA non-reporting from 2023-24
  makes NYSED charter coverage incomplete anyway). D75/D79 are absent from NYSED CA tables.
- 2020-21 (remote year), 2021-22 and 2022-23 were not in the original scope; they were
  added 2026-06-10 — see §9 (including the 2020-21 remote-attendance caveat).

## 8. Artifacts

- `extracted/SRC{2019,2022,2024,2025}_{EM,HS}_CA.csv` — raw mdb exports (gitignored dir).
- `nysed_nyc_ca_{2018-19,2023-24,2024-25}.csv` — NYC school-level All-Students rows incl.
  charters, with derived DBN and flags.
- `joined_pairs.csv` — every joined (school, year, band) with both rates/Ns/counts and diff.
- `recon_results.json` — full per-panel stats incl. worst-10 lists and BEDS exceptions.
- `stats.json` — machine-readable key numbers.
- `ours_grade_aggregates.csv`, `ours_gradespan.csv`, `ours_school_names.csv` — DB exports.
- `inspect_csvs.py`, `reconcile.py` — the pipeline (reads only; DB access was SELECT-only).

## 9. Expanded panels (2026-06-10)

Six panels added: **2020-21, 2021-22, 2022-23** × EM/HS. Same method as §1–§2, with two
mechanical differences: the NYSED side reads the **committed** NYC-filtered CSVs at
`data/sources/nysed/SRC{2022,2024}_{EM,HS}_CA_nyc.csv` (2020-21 = SRC2022 YEAR=2021;
2021-22 = SRC2022 YEAR=2022; 2022-23 = SRC2024 YEAR=2023 — its prior-year rows), and the
DOE side is a fresh `school_year_metrics` export (verified row-identical to the original
run's export; still the single data_load `644077af`). The reimplementation was
regression-checked: it reproduces **all six original panels field-for-field** against the
committed `stats.json` before computing anything new.

Not panelable: **2017-18** is NYSED-only — SRC2019 carries YEAR=2018, but the DOE EOY
file starts at 2018-19, so there is nothing to join. **2019-20** was never published
(COVID waiver). No inter-release revision check is possible for the added years: each is
held in exactly one downloaded release (SRC2021/SRC2023 were not retrieved). BEDS→DBN:
zero structural exceptions and zero DBN collisions in all six added panels; D75/D79 rows
remain entirely absent.

### Coverage (same layout as §3)

| Panel | NYC schools | charters | district | district w/ rate | suppressed `s` | DATA_REP_FLAG=N |
|---|---|---|---|---|---|---|
| 2020-21 EM (SRC2022 Y=2021) | 1,396 | 229 | 1,167 | 1,131 | 36 | 0 |
| 2020-21 HS | 615 | 80 | 535 | 468 | 67 | 0 |
| 2021-22 EM (SRC2022 Y=2022) | 1,401 | 239 | 1,162 | 1,132 | 30 | 4 |
| 2021-22 HS | 605 | 81 | 524 | 469 | 55 | 0 |
| 2022-23 EM (SRC2024 Y=2023) | 1,402 | 244 | 1,158 | 1,131 | 27 | 6 |
| 2022-23 HS | 597 | 93 | 504 | 469 | 35 | 0 |

(Success Academy reported in all three added years — 0 of 20/29/31 EM schools flagged N —
consistent with its non-reporting starting only in 2023-24.)

### Results (same layout as §4)

| Panel | joined | NYSED-only | ours-only | mean | median | SD | p10 | p90 | ≤0.5pp | ≤1.0pp | ≤1.5pp | corr | enr ratio med [p10,p90] |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 2020-21 EM | 1,128 | 3 | 1 | −0.22 | −0.13 | 1.46 | −1.32 | +0.86 | 56% | 77% | 87% | 0.996 | 1.021 [1.000, 1.100] |
| **2020-21 HS** | 414 | 54 | 2 | **−2.16** | −0.26 | **6.38** | −7.11 | +0.29 | 51% | 64% | 67% | 0.944 | 1.000 [0.993, 1.002] |
| 2021-22 EM | 1,127 | 5 | 1 | +0.08 | +0.09 | 1.50 | −0.84 | +1.35 | 54% | 76% | 85% | 0.995 | 1.022 [1.000, 1.103] |
| **2021-22 HS** | 415 | 54 | 0 | **−4.37** | −0.15 | **7.16** | −16.45 | +0.03 | 59% | 63% | 66% | 0.915 | 1.000 [0.998, 1.001] |
| 2022-23 EM | 1,123 | 8 | 4 | −0.01 | +0.13 | 2.15 | −0.89 | +1.24 | 56% | 77% | 86% | 0.987 | 1.021 [1.000, 1.098] |
| **2022-23 HS** | 414 | 55 | 1 | **−7.32** | **−0.60** | **12.54** | −31.21 | +0.81 | 29% | 47% | 53% | **0.779** | 1.000 [0.998, 1.003] |

Enrollment-weighted aggregates over the joined sets:

| Panel | NYSED (count/enr = rate) | Ours (est/N = rate) | diff |
|---|---|---|---|
| 2020-21 EM | 125,647/505,317 = 24.86% | 125,399/510,428 = 24.57% | −0.30pp |
| 2020-21 HS | 99,430/272,815 = 36.45% | 92,947/270,656 = 34.34% | −2.10pp |
| 2021-22 EM | 166,841/480,143 = 34.75% | 169,243/484,850 = 34.91% | +0.16pp |
| 2021-22 HS | 128,146/271,973 = 47.12% | 117,546/268,867 = 43.72% | −3.40pp |
| 2022-23 EM | 148,543/474,259 = 31.32% | 150,348/477,803 = 31.47% | +0.15pp |
| 2022-23 HS | 111,833/269,766 = 41.46% | 97,233/266,148 = 36.53% | −4.92pp |

### The COVID-era HS tail (verified facts)

The EM band reconciles in all three added years on the same footing as 2018-19/2023-24
(medians −0.13/+0.09/+0.13pp; 76–77% within ±1pp). The HS band does **not** reconcile in
the tails: a one-sided population of high schools sits far **below** NYSED on
near-identical denominators.

1. Schools with our rate more than 5pp below NYSED's, per HS panel: **62 (2020-21),
   112 (2021-22), 129 (2022-23)** of ~414 joined — vs 0 in 2018-19, 0 in 2023-24 and 3 in
   2024-25. The median HS school still nearly agrees (−0.26/−0.15/−0.60pp), so this is a
   tail phenomenon, not a level shift.
2. Numerator-side: HS denominators are byte-identical at 174/298/272 of ~414 schools, and
   the worst cases have equal N with hugely different counts (e.g. 2022-23 03M291 West End
   Secondary: NYSED 318/402 = 79.1% vs ours 104/401 = 25.9%). Extreme 2020-21 examples:
   02M507 NYSED **99.8%** vs DOE 15.3%; 14K685 NYSED **100.0%** vs DOE 32.7% — identical N.
3. Publisher-level, not a reconstruction artifact: at tail schools DOE's **own published
   school-total CA matches our per-grade reconstruction exactly** (13K430 2021-22: DOE
   20.147% = ours; 02M507 2020-21: DOE 15.317% = ours; 25Q425 2022-23: DOE 36.496% = ours).
4. Same-school persistence: 86 schools are in the <−5pp tail in **both** 2021-22 and
   2022-23; only 11 overlap between 2020-21 and 2021-22 (and 10 across all three).
5. The 2022-23 EM tail (20 schools <−5pp) consists **entirely** of secondary schools that
   also appear in the HS panel (6-12/MS-HS buildings: 30Q580, 03M415, 10X141, 28Q284…) —
   the EM band inherits the HS-grade phenomenon at secondary schools; elementary-only
   schools are unaffected.
6. Brooklyn Tech illustrates entry/exit: diff +0.11 (2018-19), −0.02 (2020-21), **−12.65
   (2021-22)**, **−9.49 (2022-23)**, +0.21 (2023-24), −2.60 (2024-25).

**Cause not established from data on hand.** The shape (bimodal — exact agreement or a
massive NYSED-higher gap, school-by-school, HS grades only, 2020-21 through 2022-23) is
what a school-level reporting failure into SIRS would look like (the SIRS manual's
default-to-absent rule for missing daily-attendance records is the one documented
mechanism that inflates NYSED's numerator only — see NARROWING.md), but no document on
hand shows that this occurred or which schools it hit. Do not assert a cause.

### 2020-21 caveat (applies to both 2020-21 panels, whatever the agreement)

2020-21 was the remote/hybrid year: DOE counted remote and blended "interactions" as
attendance, and NYSED published 2020-21 as an **informational-only** year under the
federal accountability waiver. The panels happen to agree at the median (EM −0.13pp,
weighted −0.30pp), but both 2020-21 panels are flagged `caveat` in `stats.json`: the two
publishers were measuring attendance under non-comparable instructional conditions, so
agreement or disagreement that year is weaker evidence than in any other panel.

### EM trend — no precursor to the 2024-25 EM divergence

Median EM diff (ours − NYSED) in time order: **+0.39 (2018-19), −0.13 (2020-21),
+0.09 (2021-22), +0.13 (2022-23), +0.38 (2023-24), −1.84 (2024-25)**. The 2024-25 EM
divergence appears abruptly with no precursor, and its shape is different in kind:
broad-based (474 of 1,115 schools below −2pp; elementary-only schools shift as much as
secondary) vs the COVID-era EM panels (51–61 schools below −2pp, deep <−5pp tail
concentrated at secondary buildings, medians unmoved).

### Hand-verified cases (continuing §5 numbering)

**Case 4 — 01M015 PS 15 Roberto Clemente, 2020-21 EM.**
NYSED (SRC2022, YEAR=2021): ENROLLMENT 146, ABSENT_COUNT 56, ABSENT_RATE 38.4.
Ours (grades 1–5): 32×53.125000% = 17; 27×51.851852% = 14; 32×25.000000% = 8;
32×40.625000% = 13; 29×24.137932% = 7 → 59 absent / 152 = **38.82%** vs NYSED
**38.4%** (diff +0.42pp; N 152 vs 146, ratio 1.041). Reconstructed counts are integers.

**Case 5 — 13K430 Brooklyn Technical HS, 2021-22 HS (a tail school).**
NYSED (SRC2022, YEAR=2022): 5,977 / 1,963 / 32.8.
Ours: g9 1,590×6.918239% = 110; g10 1,374×13.173217% = 181; g11 1,528×18.848167% = 288;
g12 1,484×42.115902% = 625 → 1,204 / 5,976 = **20.15%** vs NYSED **32.8%** (−12.65pp on a
denominator that differs by one student). DOE's own published school-total CA for 13K430
2021-22 is 20.14725% — matches the reconstruction exactly. Same signature as the 2024-25
Brooklyn Tech case (§5 Case 2): identical universe, conflicting numerator.

**Case 6 — 25Q425 John Bowne HS, 2022-23 HS.**
NYSED (SRC2024, YEAR=2023): 3,119 / 1,298 / 41.6.
Ours: g9 1,039×33.686237% = 350; g10 823×36.816525% = 303; g11 659×34.294384% = 226;
g12 630×43.015873% = 271 → 1,150 / 3,151 = **36.50%** vs NYSED **41.6%** (−5.10pp).
DOE's published school total 36.49635% matches the reconstruction exactly. (The same
school is the single worst 2024-25 HS case at −8.7pp.)

### Artifacts (expansion)

- `docs/qa_reports/nysed/stats.json` — the six new panels appended under `panels` (same
  schema; 2020-21 panels carry a `caveat` field), plus `expansion_2026_06_10` notes.
- Scratch (gitignored, alongside §8 artifacts): `expand/expand_recon.py` (pipeline incl.
  the §4-panel regression test), `expand/ours_grade_aggregates_all.csv` (fresh DB export),
  `expand/new_panels.json`, `expand/joined_pairs_expand.csv` (all 12 panels' joined rows).